Nemotron 3 Nano Omni 30B A3B Reasoning vs Voxtral Mini 4B Realtime 2602

Nemotron 3 Nano Omni 30B A3B Reasoning

A local multimodal cousin for video with speech, long documents, charts, screens and English transcription, with tool calling, a reasoning switch and a 262,144-token context. It needs about 27 GB for NVIDIA's NVFP4 build on Blackwell, or 42 GB for FP8. The main risks are the remote code it loads and a repository where nobody answers questions.

Voxtral Mini 4B Realtime 2602

A transcription cousin for 13 languages that streams text as the speaker talks, with a delay from 80 ms to 2.4 s, under Apache 2.0 and on Mistral's API. It needs a 16 GB GPU at BF16 through vLLM, Transformers or llama.cpp, or under 4 GB at Q4_K. The main gaps are no timestamps, safety results or answers from Mistral.

Nemotron 3 Nano Omni 30B A3B Reasoning

1. Start it with vllm serve nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 --max-model-len 131072 --trust-remote-code --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --kv-cache-dtype fp8 on vLLM 0.20.0 or later, after pip install "vllm[audio]" if any audio will be sent
2. Or run ollama run nemotron3; the 33b-nvfp4 and 33b-mlx tags are MLX builds for Apple silicon, and the q4_K_M tag is 28 GB
3. Set --media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}' at launch, since vLLM otherwise samples about 32 frames per video
4. Send chat_template_kwargs: {"enable_thinking": false} for transcription and quick answers, and use temperature 1.0 with top_k 1 for ASR as the card recommends
5. Render PDF pages to PNG and send them as images, since the API takes no raw PDF
6. Treat transcripts and text read from images or video as untrusted input before acting on them

Voxtral Mini 4B Realtime 2602

1. Serve it with VLLM_DISABLE_COMPILE_CACHE=1 vllm serve mistralai/Voxtral-Mini-4B-Realtime-2602 --compilation_config '{"cudagraph_mode": "PIECEWISE"}' and install soxr, librosa and soundfile; mistral_common 1.9.0 or later comes with vLLM
2. Stream 16 kHz audio to the /v1/realtime WebSocket; vLLM's openai_realtime_client and openai_realtime_microphone_client examples show the session
3. Set temperature 0.0. Change the delay with transcription_delay_ms in tekken.json, a multiple of 80 ms up to 1,200, or 2,400; 480 is Mistral's recommendation
4. One text token is 80 ms, so set --max-model-len to at least seconds / 0.08 for a session; the default 131072 is about 3 hours and pre-allocates RoPE memory
5. In Transformers (5.2.0 or later), resample with processor.feature_extractor.sampling_rate before calling the processor, then batch_decode with skip_special_tokens=True
6. For a small machine, ggml-org's Q4_K_M GGUF with its mmproj runs in llama.cpp, and Mistral Experimental's Q4_K runs in audio.cpp
7. Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions